Predictive remapping and allocentric coding as consequences of energy efficiency in recurrent neural network models...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Stable visual perception during eye movements relies on predictive computations using efference copies.
• The neural basis for these complex computations and their underlying connectivity is not fully understood.
• This study explores whether simple physical principles, like energy efficiency, can explain these phenomena.

Purpose of the Study:

• To investigate if energy efficiency constraints can drive the emergence of predictive remapping and allocentric coding.
• To model how neural networks might develop complex visual processing mechanisms.
• To understand the potential origins of sophisticated neural computations from basic principles.

Main Methods:

• A recurrent neural network was trained on sequences of fixation patches and saccadic efference copies.
• The model was optimized to minimize energy consumption, specifically preactivation.
• The network's emergent computational properties, including reference frame coding, were analyzed.

Main Results:

• Energy-efficiency optimization alone led to the emergence of targeted inhibitory predictive remapping.
• The model developed the ability to re-code egocentric eye coordinates into an allocentric reference frame.
• Predictive remapping in the model was dependent on the learned allocentric coding.

Conclusions:

• Complex neural computations like predictive remapping and allocentric coding can emerge from energy-efficiency constraints.
• Simple physical principles may be sufficient to explain sophisticated neural mechanisms for visual stability.
• This work provides insights into how biological neural networks might evolve complex functions.